Daily activity patterns from wearable accelerometry predict physical frailty and concern about falling

Abstract

Physical frailty and concern about falling are interrelated geriatric conditions that significantly impair mobility, limit daily activity, and reduce life quality among older adults. Traditional methods for assessing frailty and fall concern rely heavily on questionnaires such as the Fried Frailty Phenotype (FFP) and the Falls Efficacy Scale-International (FES-I). These tools, although clinically established, are subjective, require professional administration, and cannot capture the mutual influence between the two conditions. In this study, we propose a wearable-sensor-based, objective framework that simultaneously predicts physical frailty status and level of concern about falling by analyzing real-world activity patterns. We collected continuous chest acceleration data from 146 participants over a 48-hour period using a pendant sensor. These activity sequences were transformed into barcodes and analyzed using global complexity metrics (e.g., single- and multi-scale entropy) and local sequential dynamics via bidirectional LSTM networks. A multi-task deep learning model with attention mechanisms was proposed to jointly predict frailty and fall concern. Our model achieved high predictive performance for both tasks, achieving F1 scores of 93.12% for physical frailty and 86.27% for concern about falling. This study provides the first joint modeling of physical frailty and concern about falling using long-term real-world accelerometry data.

Introduction

Physical frailty and concern about falling are two prevalent expressions of population aging. Physical frailty refers to a medical syndrome characterized by diminished strength, endurance, and reduced physiological function, which increases an individual’s vulnerability to developing greater dependency and/or mortality¹. This condition results from the cumulative decline of multiple physiological systems, including the musculoskeletal, cardiovascular, metabolic, immune, neurological, and endocrine systems². Concern about falling is a psychological construct reflecting apprehension about experiencing a fall. While mild concern may promote safety among robust older adults, elevated concern about falling frequently becomes maladaptive in frail individuals, leading to excessive activity restriction, social withdrawal, and progressive physical deterioration^3–5. A well-documented bidirectional relationship exists between frailty and concern about falling. Elevated concern about falling increases the likelihood of pre-frailty and frailty onset, while individuals with diminished physiological capacity are more prone to fear-related behavior patterns^6,7. Breaking this downward spiral requires early and accurate detection of both physical and psychological vulnerabilities. Timely identification not only enables personalized rehabilitation and fall prevention strategies but also helps clinicians intervene before irreversible decline occurs. Critically, understanding how frailty and fall-related concern co-develop can support nuanced risk stratification and effective, individualized interventions.

Traditional measurements for physical frailty and fall concerns often rely on questionnaires, scales, and simple experimental tests. The Fried frailty phenotype (FFP)⁸ is a well-established method for identifying frailty by evaluating symptoms and signs associated with biological aging. It assesses the older adults’ performance based on five criteria: unintentional weight loss, self-reported exhaustion, low physical activity, reduced muscle strength, and slow walking speed. Accordingly, individuals are categorized as robust (no criteria met), pre-frail (one or two criteria), or frail (three or more criteria). The Falls Efficacy Scale–International (FES-I) is an established and extensively utilized instrument in both clinical and research settings for assessing an individual’s level of concern about falling (University of Manchester, https://documents.manchester.ac.uk/display.aspx?DocID=38565). The FES-I is a brief and easy-to-understand tool measuring an individual’s concern about falling during 16 social and physical activities, both indoors and outdoors. Based on the total score, individuals are classified into high- or low-concern about falling groups. However, these tools require professional supervision and are limited in their applicability to telehealth or remote monitoring settings. Self-report questionnaires are also susceptible to personal biases, as respondents may interpret items inconsistently or provide socially desirable or emotionally influenced answers^9,10. To overcome these limitations, there is a critical need for a comprehensive, long-duration, questionnaire-free framework that allows for the simultaneous analysis of both conditions and facilitates the discovery of their interrelationship.

Wearable accelerometry has emerged as a promising modality for quantifying real-world activity patterns and inferring underlying health status for older adults. Triaxial accelerometer data can reveal activity patterns and intensity levels, supporting the assessment of key geriatric conditions such as physical frailty and concern about falling. Prior studies have demonstrated that wearable sensors provide valuable insights into frailty-related physical condition among vulnerable older adults, with gait and activity features derived from triaxial accelerometry serving as effective indicators of pre-frailty and frailty status¹¹. Related work combining accelerometer-derived gait characteristics, demographic variables, and nonlinear complexity measures has reported good to strong performance for binary frailty classification, with accuracies typically exceeding 80% in laboratory-based or mixed laboratory and free-living settings^12,13. Pendant-based wearable devices have also been used to derive digital biomarkers of frailty, showing promise in predicting chemotherapy resilience among veterans with cancer¹⁴. For concern about falling, wearable-derived mobility and activity behavior patterns have likewise been shown to be informative. Previous studies reported that higher concern about falling, as measured by the FES-I, is linked to reduced diversity and temporal complexity of free-living physical activity patterns extracted from wearable accelerometry¹⁵, as well as to altered real-world mobility performance captured by wearable sensors¹⁶. Beyond correlation-based analyses, classification-oriented approaches have applied multi-time-scale topic modeling to triaxial acceleration signals collected in free-living conditions, achieving moderate classification performance (e.g., accuracy around 70%) for distinguishing between high and low concern about falling¹⁷. Despite prior efforts to assess physical frailty and concern about falling separately using wearable data, most existing approaches remain limited by their dependence on handcrafted, static features, such as step count, gait speed, or activity duration, extracted through conventional feature engineering. These features often oversimplify the rich temporal dynamics and multiscale variability present in continuous accelerometry signals. As a result, they fail to capture subtle fluctuations in behavior that are critical for identifying early or overlapping manifestations of frailty and fall concern. Moreover, nearly all existing models treat these conditions independently, overlooking their known interaction. This neglect of joint modeling further restricts predictive accuracy and prevents a more integrated understanding of aging-related vulnerability.

In this study, we propose a novel framework to jointly assess physical frailty and concern about falling in older adults by modeling their daily physical activity patterns derived from wearable triaxial accelerometry. Continuous 48 h acceleration data were segmented into 16 physical activity states based on intensity and duration, then visualized as barcodes to reveal temporal structure. From these sequences, we extracted global features, including single-scale and multi-scale entropy measures as complexity indicators, to capture long-term behavioral trends. Meanwhile, we employed a bidirectional Long Short-Term Memory (LSTM) network to learn local temporal features, reflecting short-term fluctuations and activity transitions. A multi-task deep learning neural network, augmented with an attention mechanism, was further developed to perform dual classification tasks: physical frailty and concern about falling. The model was then rigorously evaluated through standard predictive metrics and ablation analyses, while attention-weight heatmaps revealed task-specific feature contributions and the shared behavioral signatures underlying both conditions. By leveraging real-world wearable data and interpretable multi-task learning, this study provides a scalable, objective, and behaviorally grounded approach for early risk stratification in community-dwelling older adults, offering insights for future personalized interventions.

Results

Key demographic and psychological factors of the 146 participants who completed the 48-hour continuous data collection phase for each condition are summarized in Tables 1 and 2. The collected information included age, sex, body mass index (BMI), scores from the Center for Epidemiological Studies Depression (CES-D), and Mini-Mental State Examination (MMSE). The cohort was stratified into subgroups based on two established measures: FES-I and FFP, with the distribution across their combinations provided in Suppl. Table S1. According to the FES-I, participants were divided into two groups: 63 participants with low concern about falling and 83 participants with high concern about falling. Based on the FFP, participants were categorized into three groups: 50 robust, 73 pre-frail, and 23 frail individuals. Significant differences were observed in age, BMI, CES-D, MMSE, and FES-I scores between the high- and low-concern groups. Similarly, regarding physical frailty, age, BMI, CES-D, and FES-I scores significantly varied across the three frailty classifications. Notably, the FES-I scores increased progressively from robust to frail groups, indicating that higher levels of frailty were accompanied by greater concern about falling. This trend was further supported by a Chi-square test, which demonstrated a significant association between physical frailty and concern about falling (p < 0.001), reinforcing the bidirectional relationship between the two conditions.

Table 1.

Demographic and psychological characteristics of participants with different levels of physical frailty (robust, pre-frail, frail)

	Robust	Pre-frail	Frail	p-value
Number	50	73	23
Age	74.370 ± 6.570	79.150 ± 8.140	81.570 ± 8.610	< 0.001
Female, n (%)	42(84)	53(72)	18(78)	0.330
BMI	25.930 ± 4.430	29.290 ± 7.240	30.160 ± 5.290	0.003
CES-D	6.800 ± 5.657	7.255 ± 6.677	12.68 ± 6.841	< 0.001
MMSE	29.103 ± 1.123	28.536 ± 1.573	28.647 ± 1.733	0.106
FES-I	18.370 ± 1.080	25.570 ± 2.500	33.190 ± 8.140	< 0.001

Values are presented as mean ± s.d. (range) unless otherwise indicated. Categorical variables are presented as n (%). BMI Body Mass Index, CES-D Center for Epidemiologic Studies Depression Scale, MMSE Mini-Mental State Examination, FES-I Falls Efficacy Scale – International. The FFP categorizes individuals into three groups: robust (N = 50; meeting 0 of 5 criteria), pre-frail (N = 73; meeting 1–2 of 5 criteria), and frail (N = 23; meeting ≥ 3 of 5 criteria). Bold values indicate statistically significant p-values (p < 0.05).

Table 2.

Demographic and psychological characteristics of participants with different levels of concern about falling (low, high)

		Low concern about falling	High concern about falling	p-value
Number		63	83
Age		76.330 ± 7.150	79.220 ± 8.757	0.010
Female, n (%)		50(79)	63(75)	0.767
BMI		26.912 ± 5.354	29.178 ± 6.552	0.024
CES-D		5.929 ± 5.675	9.450 ± 6.952	0.045
MMSE		29.14 ± 1.120	28.433 ± 1.643	0.054
FES-I		19.050 ± 1.780	25.990 ± 7.260	< 0.001
FFP, n (%)	Robust	35(56)	15(18)	< 0.001
	Pre-frail	26(41)	47(57)
	Frail	2(3)	21(25)

Values are presented as mean ± s.d. (range) unless otherwise indicated. Categorical variables are presented as n (%). BMI Body Mass Index, CES-D Center for Epidemiologic Studies Depression Scale, MMSE Mini-Mental State Examination, FES-I Falls Efficacy Scale – International, FFP Fried Frailty Phenotype. A FES-I score ≥ 23 indicates high concern about falling, while a score < 23 indicates low concern. A FES-I score ≥ 23 indicates high concern about falling, while a score < 23 indicates low concern. Bold values indicate statistically significant p-values (p < 0.05).

Three-dimensional axial acceleration data were collected from 146 participants: vertical (z-axis), horizontal (x-axis), and lateral (y-axis). Signal magnitude vectors (SMV) were computed to quantify overall movement intensity by integrating accelerations across all three axes. Using predefined thresholds for intensity and duration on each axis, wearable sensor data were classified into 16 distinct physical activity (PA) states, capturing a wide range of behaviors including lying, sitting, and walking. Specifically, raw sensor data were transformed into per-second sequences of PA states. These sequences were then encoded as color-coded PA barcodes, providing a visual representation of activity distribution across the 48 h monitoring window. PA states and barcodes stratified by levels of physical frailty and concern about falling reveal distinct group-specific activity patterns (Fig. 1). Notably, individuals with high concern about falling showed reduced color diversity and fewer state transitions, indicating lower variability in activity patterns. Robust participants exhibited more warm-toned barcodes, reflecting frequent high-intensity activity, while frail individuals displayed predominantly cool-toned patterns, corresponding to low activity levels and limited mobility over the 48 h period.

Fig. 1. PA state distributions and barcode representations for participants stratified by frailty status and concern about falling.

Subgroups: a Robust with low concern about falling. b Robust with high concern about falling. c Pre-frail with low concern about falling. d Pre-frail with high concern about falling. e Frail with low concern about falling. f Frail with high concern about falling.

In this study, features were categorized into global and local representations to capture distinct aspects of physical activity patterns. Global features represent high-level statistical attributes, reflecting broad behavioral trends over extended periods, such as the analyzed 48 h sequences. In contrast, local features emphasize fine-grained details extracted from shorter time windows to highlight moment-to-moment variations such as transitions between physical activity states. As demonstrated in Suppl. Table S2, this hourly representation provided superior predictive performance compared with finer temporal resolutions. These subtle shifts provide valuable insights into short-term activity fluctuations in behavior¹⁸. As part of the analysis, a total of 206 global features were extracted, which included single-scale and multi-scale complexity measures. Single-scale complexity measures included activity percentage, sample entropy (SE), information entropy (Hn), and Lempel-Ziv complexity (LZC). Multi-scale features consisted of multi-scale sample entropy (MSE), multi-scale information entropy (MHn), scaled from 1 to 100, along with their average values (average MSE and average MHn).

Following statistical analysis, 48 features were found to be significant (p ≤ 0.05) between groups for at least one of the two conditions: physical frailty and fall concern. This subset included demographic and psychological (age, BMI, CES-D), and physiological complexity features such as SE, Hn, LZC, activity percentage, average MSE, average MHn, and 39 individual MHn scales. Tables 3 and 4 summarize the significance analyses of the nine selected global features, excluding the 100 time-scale MSE and MHn features. The full statistical results for all 100 individual MSE and MHn features are available in Supplementary Table S3. Additionally, 16 local variables were derived from each participant’s 48 h PA state sequences using an LSTM-based temporal neural network, enabling the capture of detailed motion dynamics and sequential activity patterns.

Table 3.

Statistical significance analysis of selected features across physical frailty groups (robust, pre-frail, frail)

	Robust	Pre-frail	Frail	p-value
CES-D	6.800 ± 5.657	7.255 ± 6.677	12.68 ± 6.841	0.002
Age	74.370 ± 6.570	79.150 ± 8.140	81.570 ± 8.610	< 0.001
BMI	25.930 ± 4.430	29.290 ± 7.240	30.160 ± 5.290	0.007
Average MHn	5.101 ± 3.035	4.505 ± 2.661	4.426 ± 1.865	0.448
Average MSE	7.323 ± 4.029	3.393 ± 1.698	0.155 ± 0.617	< 0.001
Hn	0.573 ± 0.109	0.557 ± 0.112	0.566 ± 0.192	0.807
SE	0.078 ± 0.293	0.040 ± 0.018	0.019 ± 0.008	< 0.001
LZC	0.029 ± 0.011	0.027 ± 0.020	0.024 ± 0.020	0.679
%activity	0.131 ± 0.110	0.127 ± 0.150	0.095 ± 0.130	0.002

Values are presented as mean ± s.d. (range) unless otherwise indicated. Group sizes were: robust (n = 50), pre-frail (n = 73), and frail (n = 23). CES-D Center for Epidemiologic Studies Depression Scale, BMI Body Mass Index, Average MHn Multi-scale information entropy averaged over all considered time scales (1-100), Average MSE Multi-scale sample entropy averaged over all considered time scales (1–100), Hn Information entropy, SE Sample entropy, LZC Lempel-Ziv complexity, %activity Dynamic activity percentage. Bold values indicate statistically significant p-values (p < 0.05).

Table 4.

Statistical significance analysis of selected features across concern about falling groups (low vs. high)

	Low concern about falling	High concern about falling	p-value
CES-D	5.929 ± 5.675	9.450 ± 6.952	< 0.001
Age	76.330 ± 7.150	79.220 ± 8.757	0.064
BMI	26.912 ± 5.354	29.178 ± 6.552	0.038
Average MHn	4.029 ± 0.903	3.117 ± 0.746	0.041
Average MSE	5.979 ± 3.862	2.781 ± 1.626	< 0.001
Hn	0.567 ± 0.104	0.412 ± 0.089	0.048
SE	0.074 ± 0.030	0.033 ± 0.017	< 0.001
LZC	0.023 ± 0.095	0.019 ± 0.010	0.018
%activity	0.150 ± 0.150	0.084 ± 0.082	0.021

Values are presented as mean ± s.d. (range) unless otherwise indicated. Group sizes were: low concern about falling (n = 63) and high concern about falling (n = 83). CES-D Center for Epidemiologic Studies Depression Scale, BMI Body Mass Index, Average MHn Multi-scale information entropy averaged over all considered time scales (1–100), Average MSE Multi-scale sample entropy averaged over all considered time scales (1–100), Hn Information entropy, SE Sample entropy, LZC Lempel-Ziv complexity, %activity Dynamic activity percentage. Bold values indicate statistically significant p-values (p < 0.05).

The performance of the proposed multi-task learning model was compared with five single-task learning models (Table 5): logistic regression (LR), decision tree (DT), support vector machine (SVM), LSTM, and Time Series Transformer (TST). For the physical frailty classification task, SVM achieved the highest accuracy among the single-task models (77.45%), followed by LR (74.49%) and DT (70.98%), whereas the sequence-based models showed lower performance (LSTM: 50.00%, TST: 59.09%). For the concern about falling classification task, DT achieved the highest accuracy (80.45%) and F1-score (79.74%), while LR and SVM showed comparable performance (both 79.27% accuracy). LSTM and TST showed relatively lower performance across all evaluation metrics (accuracy: 47.73% and 54.55%, respectively). In contrast, the multi-task learning model with attention layers outperformed all single-task models, achieving higher values across all evaluation metrics, including accuracy, precision, recall, and F1 score for both tasks. Specifically, the model reached an accuracy of 93.18%, precision of 93.17%, recall of 93.18%, and F1-score of 93.12% for physical frailty classification. For concern about falling, it achieved an accuracy of 84.09%, precision of 84.62%, recall of 88.00%, and F1-score of 86.27%. The confusion matrices for both classification tasks provide further insight into model performance (Fig. 2). The hyperparameters used for single-task and multi-task models are detailed in Suppl. Tables S4 and S5, respectively.

Table 5.

Performance evaluation of single-task and multi-task learning model for classifying physical frailty and concern about falling

		Physical Frailty				Concern About Falling
		Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)
Single-task Learning Models	LR	74.49	80.51	72.91	73.62	79.27	81.20	75.60	76.48
	SVM	77.45	83.98	75.83	76.84	79.27	82.82	74.47	75.57
	DT	70.98	73.39	71.97	71.35	80.45	80.98	80.87	79.74
	LSTM	50.00	25.00	50.00	54.16	47.73	48.05	47.73	47.86
	TST	59.09	65.73	59.09	55.67	54.55	53.99	54.55	54.16
Multi-task Learning Model (MTL)		93.18	93.17	93.18	93.12	84.09	84.62	88.00	86.27

LR Logistic regression, SVM Support vector machine, DT Decision tree, LSTM Long short-term-memory network, TST Time series transformer,

MTL Multi-task learning model.

Within the multi-task model architecture, physical frailty is designated as the main task and concern about falling as the auxiliary task.

Fig. 2. Confusion matrices for multi-task predictions.

a Physical frailty b Concern about falling.

To evaluate the contribution of each model component, we conducted ablation experiments by individually removing the demographic/psychological, global, local, and attention modules (Table 6). For physical frailty, the full MTL model achieved balanced and consistently strong performance across all metrics. The most pronounced degradation occurred when global features were removed, with a sharp drop in precision (55.00%) and F1-score (55.48%), indicating severe loss of discriminative capacity. Removing the attention mechanism also substantially reduced recall (65.91%) and F1-score (67.81%), suggesting impaired integration of heterogeneous features. Excluding local temporal features mainly affected recall (77.27%) and F1-score (77.67%), underscoring the importance of short-term dynamics, while removing demographic and psychological variables led to a moderate but consistent decline across metrics (recall = 84.09%, F1 = 84.08%). For concern about falling, removing global features caused the largest decline, particularly in precision (61.54%) and accuracy (56.82%), demonstrating their dominant role. Excluding the attention module reduced both precision (75.00%) and recall (72.00%), while removing local features preserved high recall (88.00%) but lowered precision (78.57%). Demographic and psychological features provided supportive but secondary gains (F1 = 82.35%). Overall, these results demonstrate that global, local, demographic/psychological, and attention-based multitask components contribute complementary and synergistic benefits to overall model performance.

Table 6.

Ablation analysis of the multi-task learning architecture

	Physical Frailty				Concern about Falling
	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)
MTL	93.18	93.17	93.18	93.12	84.09	84.62	88.00	86.27
MTL-demo. & psy.	84.09	84.22	84.09	84.08	79.55	80.77	84.00	82.35
MTL-global	63.64	55.00	63.64	55.48	56.82	61.54	64.00	62.75
MTL-local	77.27	81.20	77.27	77.67	79.55	78.57	88.00	83.02
MTL-attention	65.91	74.24	65.91	67.81	70.45	75.00	72.00	73.47

MTL Multi-task learning model.

MTL – demo. & psy. Multi-task learning model with demographic and psychological feature components removed.

MTL – global Multi-task learning model with global feature component removed.

MTL – local Multi-task learning model with local feature component removed.

MTL – attention Multi-task learning model with attention mechanism component removed.

Feature contribution visualization

To investigate the role of each feature within the attention mechanism, we visualized attention heatmaps for both classification tasks (Fig. 3). These heatmaps illustrate how the model prioritizes features, highlighting their relative importance and contribution to predictive performance. Through these visualizations, we identified the most influential features for optimizing classification accuracy and gained deeper insights into the internal behavior of the attention mechanism. Our framework systematically evaluates a total of 64 features, consisting of 2 demographic, 1 psychological, 45 global, and 16 local variables, all of which undergo attention-based refinement. Specifically, features with attention weights above the 75th percentile within each task’s attention-weight distribution were selected as the most salient subset. This approach focuses on the upper tail of the distribution while accounting for task-specific differences in attention-weight scaling and avoiding excessive sparsity. In the primary task—physical frailty classification, the most salient features include psychological variable CES-D, a set of multiscale movement complexity descriptors (MHn8, MHn9, MHn13–15, MHn33, MHn34, and MHn37) as well as several local temporal features (Local4–Local8, Local13, and Local16). In the auxiliary task—concern about falling, the features receiving the highest attention weights comprised psychological and physiological variables (CES-D and BMI), global activity descriptors (LZC, activity percentage, and SE), and a broader collection of multiscale MHn measures (MHn5, MHn8–10, MHn13–15, MHn18, MHn29, MHn30, and MHn34). Notably, both tasks consistently emphasized a shared subset of multiscale MHn features (MHn8, MHn9, MHn13–16, and MHn34) and CES-D, reinforcing their relevance to both classification outcomes in predictive modeling. By averaging attention weights across samples, we quantified feature importance and gained deeper insight into the internal prioritization process of the model. The attention-weight distributions motivating the percentile-based threshold are provided in Suppl. Fig. S1.

Fig. 3. Visualization of the average attention weights illustrating the relative contribution of each input feature to the model’s prediction.

Features are grouped into four categories: demographic and psychological (e.g., age, BMI, and CES-D), global complexity measures (e.g., entropy and activity-derived metrics), and local activity patterns. Higher attention weights indicate greater importance in influencing model decisions.

Discussion

In this study, we introduced a multi-task deep learning framework to jointly classify physical frailty and concern about falling using wearable triaxial accelerometry. By combining entropy-based global features and local temporal representations derived from LSTM networks, our method enables fine-grained characterization of daily activity behavior in real-world settings. The incorporation of attention mechanisms further enhanced model interpretability by identifying key features contributing to each classification task, ultimately achieving high predictive accuracy across both conditions.

A central objective of our study was to explore the association between physical frailty and concern about falling. Statistical analysis confirmed a significant dependency between the two conditions (Tables 1 and 2), supporting previous findings that elevated concern often accompanies or precedes physical deterioration¹⁹. This reinforcing loop—wherein fear restricts activity and accelerates frailty—underscores the value of modeling the two conditions jointly. Our multi-task framework was therefore grounded in this interdependence, yielding superior classification performance compared to single-task approaches. Prior wearable-based studies have reported moderate to strong performance for assessing concern about falling and physical frailty using single-task models. For instance, multi-time-scale topic modeling of free-living accelerometry achieved approximately 71% accuracy for concern-about-falling classification¹⁷, while binary frailty classification using IMU-derived gait features typically exceeded 80% accuracy in laboratory or mixed settings^12,13. A comparative overview of these wearable-based studies is provided in Suppl. Table S6. The present work extends this literature by addressing joint modeling and a multi-class frailty setting using exclusively free-living data.

Analysis of physical activity barcodes revealed that both frailty and fall concern were associated with declines in behavioral complexity. Participants with robust physical function and low concern exhibited the most varied activity profiles, whereas increasing frailty led to more static, monotonous movement states, and increasing fear led to variability in physical activity states (Fig. 1). These trends were captured quantitatively through entropy measures, particularly MHn and MSE, which consistently ranked among the most informative features. These findings support previous work demonstrating the utility of multi-scale entropy in capturing nuanced changes in human movement¹⁵. Attention-based feature analysis (Fig. 3) further confirmed that multi-scale entropy contributed obviously to both classification tasks, reinforcing movement complexity as a key indicator of mobility-related health.

Additionally, attention-based feature analysis also revealed a divergence in the temporal dependencies of the two conditions (Fig. 3). Both concern about falling and physical frailty were predominantly influenced by global features, particularly multiscale MHn measures that characterize the complexity and temporal organization of physical activity across multiple time scales. Beyond these shared global influences, physical frailty exhibited a stronger dependence on local temporal features reflecting short-term deviations in motor control, suggesting an increased vulnerability to transient impairments in movement dynamics. This pattern likely reflects partially distinct neurobehavioral mechanisms, whereby physical frailty is more closely associated with biomechanical degradation and impaired motor control, whereas concern about falling is more strongly shaped by cognitive–affective processes influencing sustained behavioral regulation^20–23. Such differentiation may inform the development of stratified monitoring strategies and targeted interventions tailored to the specific trajectories of decline observed in aging populations. Importantly, despite the relatively small and imbalanced cohort, these feature-contribution patterns were highly consistent across all cross-validation folds. Both training dynamics and cross-fold attention distributions showed stable convergence without systematic divergence, indicating that the learned attention weights capture task-relevant and generalizable characteristics (Suppl. Fig. S1 and S2). This stability was further supported by fold-specific attention heatmaps, which exhibited largely consistent similar spatial patterns of feature weighting across all folds (Suppl. Fig. S3).

Further supporting this interpretation, demographic variables such as age, BMI, and psychological factors (particularly the CES-D) demonstrated significant predictive relevance, as confirmed by group difference analyses. Age and BMI were significantly associated with physical frailty and concern about falling, consistent with known risk factors in geriatric populations²⁴. Notably, CES-D emerged as a key predictor in both tasks, emphasizing the role of depressive symptoms in modulating physical activity and fall-related fear. This finding aligns with studies showing that fear of falling is shaped by sociodemographic, physical, and psychological health factors, reinforcing the connection between psychological distress and mobility deterioration²⁵. Moreover, the mediating role of physical frailty in the relationship between depression and fall risk suggests that emotional health can actively exacerbate physiological decline, culminating in elevated susceptibility to adverse outcomes²⁶. These results support a growing consensus that physical and mental health are tightly interwoven in late life, warranting integrated screening and intervention approaches.

While intervention design was beyond the scope of this study, our stratified findings in daily behaviors suggest that individuals with high concern but no frailty may benefit from cognitive–behavioral strategies to prevent fear-driven activity avoidance^27,28. Conversely, frail individuals with low concern may lack risk awareness, pointing to the need for targeted education and physical rehabilitation²⁹. Those exhibiting both conditions likely represent a high-risk subgroup requiring comprehensive, multidisciplinary care³⁰.

Our work effectively enhances the classification performance for both physical frailty and concern about falling, achieving high accuracy in the smallest yet most clinically relevant subgroup—individuals experiencing both high concern about falling and physical frailty. However, several limitations should be acknowledged. First, the cohort was relatively small (n = 146), with a predominance of female participants, and derived from a secondary analysis of studies focused on frailty and fall risk, which may limit generalizability and does not fully capture the dynamic progression of frailty and concern about falling; validation in larger, more diverse, and longitudinal datasets is needed. Second, physical activity was measured using a pendant-worn sensor under free-living conditions, where variations in adherence, positioning, and movement artifacts (e.g., sensor swinging) may introduce noise. Although wavelet-based filtering, robust temporal activity descriptors (e.g., cadence, duration, and sustained posture), and long-duration aggregated features were used to reduce the influence of transient perturbations, such variability cannot be fully eliminated. Future work could explore improved sensor designs, standardized wearing protocols, and pendulum-based dynamic modeling to better account for rope–sensor motion. Finally, although attention-based analysis improved interpretability, attention-guided model simplification was not explored and may further enhance efficiency for real-time or resource-constrained applications.

In conclusion, this study represents the first comprehensive investigation demonstrating that wearable triaxial accelerometry data, combined with global and local feature representations and multi-task deep learning methods, can accurately estimate two prevalent conditions – physical frailty and concern about falling in community-dwelling older adults. These findings provide a foundation for future health monitoring systems and targeted intervention strategies, offering clinical insights for aging populations.

Methods

This study presents a secondary analysis of baseline data collected from 146 older adults who participated in two longitudinal cohort studies focused on mobility, cognitive function, and aging-related outcomes in community-dwelling older adults. The methodologies and primary objectives of these original studies were detailed in a prior publication¹⁶. Participants were systematically recruited from a broad range of care and living environments—including primary, secondary, and tertiary healthcare facilities, community health providers, assisted living facilities, retirement homes, aging service organizations, and independent community residences—to ensure a diverse and representative sample across varying levels of functional ability and care needs. Eligibility criteria required participants to be 65 years of age or older, ambulatory, and capable of walking at least 1.8 meters ( ~ 6 feet) without hands-on assistance, though assistive devices such as canes were permitted. Exclusion criteria included cognitive impairment (as assessed by the MMSE or the Montreal Cognitive Assessment [MoCA]^31,32), terminal illness, acute mobility-affecting conditions (e.g., recent stroke, fracture, or foot ulcer), severe depression, active medical therapies that may influence mobility (e.g., chemotherapy), lack of capacity to provide informed consent, or unwillingness to participate. All participants provided written informed consent, and study protocols received approval from the respective Institutional Review Boards.

Participants were stratified based on two established measures. According to the FES-I, participants were categorized into two groups: low concern about falling (FES-I < 23, N = 63) and high concern (FES-I ≥ 23, N = 83). Independently, based on the FFP, participants were classified into three groups: robust (N = 50), pre-frail (N = 73), and frail (N = 23). It is important to note that, because the original studies primarily focused on frailty and fall risk, the recruitment strategy may have resulted in an overrepresentation of individuals exhibiting early signs of frailty or heightened concern about falling. Consequently, the findings from this secondary analysis may be more reflective of this higher-risk subpopulation rather than the broader older adult population. To monitor physical activity, participants wore a pendant-mounted triaxial accelerometer (PAMSys1, Biosensics, Boston, MA, USA) for 48 continuous hours. The sensor, worn around the neck, sampled data at 50 Hz within a ± 2 G range (G = 9.8 m/s²) and operated with a low current consumption of 400 µA, allowing full data collection on a single charge. Participants were instructed to wear the device at all times, except while showering, due to its lack of water resistance.

PA states were defined by combining various physical activity dimensions. An algorithm was developed to recognize PA using triaxial acceleration signals in the X, Y, and Z directions. The signal magnitude vector (SMV) was represented by the sum of the squares of these accelerations (Eq. 1):

$$\mathrm{SMV}=\sqrt{a_v^2+a_f^2+a_l^2}$$ 1

where $a_v$, $a_f$, and $a_l$ represent acceleration magnitude in the vertical, frontal, and lateral directions, respectively. Gravitational acceleration was not explicitly subtracted from the raw acceleration signals prior to SMV computation. Instead, to suppress low-frequency components—including quasi-static gravitational contributions—a five-level discrete wavelet transform (DWT) was applied to the SMV using the Daubechies-4 (db4) wavelet basis³³. This wavelet-based filtering effectively attenuates slowly varying components while preserving movement-related acceleration dynamics. Each data point of the filtered SMV signal was labeled as static or dynamic based on its temporal acceleration pattern. Static states were characterized by stable, gravity-aligned acceleration with low variability, whereas dynamic states exhibited pronounced, periodic acceleration fluctuations, particularly rhythmic oscillations indicative of locomotor activity. Static periods were identified based on sustained gravity-aligned acceleration signals ( ≈ 1 G) along posture-specific axes: vertical for sitting, and frontal or lateral for lying. Consecutive static samples were grouped into segments, and segments shorter than 30 s were regarded as transient fluctuations and merged into the preceding state. Static behaviors were subsequently subdivided according to filtered SMV fluctuation using a threshold of 0.2 m/s², resulting in two lying states (States 1–2) and two sitting states (States 3–4). Walking states were categorized based on average cadence and bout duration, resulting in 12 walking-related PA states (States 5–16). Cadence was computed within each detected walking bout from step-event timing derived from periodic vertical-axis acceleration oscillations. Step events were identified as prominent, regularly spaced initial foot contact (IC) events in the processed vertical acceleration signal during walking³⁴. Cadence was then computed at the bout level based on the timing of successive IC events (Eq. 2).

$$C=\frac{60\cdot (n-1)}{t_n-t_1}$$ 2

where $t_1$ and $t_n$ denote the timestamps (s) of the first and last detected IC events in a walking bout, $n$ is the number of detected steps, and 60 converts cadence to steps/min. Cadence was grouped into four ranges ( < 50, 50–80, 80–140, and > 140 steps/min), and bout duration was categorized using thresholds of 30 and 120 s, with short walking bouts retained as valid dynamic segments to capture fragmented free-living mobility patterns³⁵. All samples within a given walking bout were assigned the same PA-state label. Consequently, over a 48 h period, each individual’s motion data formed a sequence of around 172,800 PA states. This sequence was represented as a color-coded temporal “barcode”, serving as a time-series encoding and visualization of long-term physical activity patterns. Specifically, the continuous PA-state sequence was mapped to a temporal representation in which each color corresponds to a discrete activity state. Cooler tones represent lower-intensity states, while warmer tones indicate higher-intensity states (Fig. 4).

Fig. 4. Illustration of the 16 Physical Activity (PA) states and corresponding barcode representation.

a Definition of each PA state. b Example of PA states and barcode encoding.

This study investigated a diverse range of features derived from individuals’ demographic and psychological attributes and PA state sequences. Statistical analysis was then applied to identify the most significant features for further modeling.

To quantify the intensity of daily activity patterns, both single-scale and multi-scale complexity indices were derived from PA state sequences. Feature extraction was guided by established methods in time-series and physiological signal analysis, particularly entropy- and complexity-based measures that have been widely used to characterize variability and dynamical structure in physical activity data^15,17. Single-scale features encompassed activity percentage, Lempel-Ziv complexity (LZC) (Eq. 3), SE (Eq. 4), and Hn (Eq. 5). Multi-scale features included MSE (Eq. 6) and MHn (Eq. 7). LZC is a measure of the number of new patterns that emerge as a sequence develops³⁶. SE reflects time series complexity and is widely applied in fault diagnosis and biological time series analysis, with higher values indicating greater complexity³⁷. Hn measures the average information content after redundancy elimination, with lower values indicating reduced complexity and greater self-similarity³⁸. MSE evaluates signal complexity by integrating multiscale and nonlinear characteristics. Through a coarse-graining process, the signal generates multiple sub-signals across scales, calculating sample entropy for each scale from 1 to 100, thereby enriching the interpretation of entropy and offering a detailed view of signal properties³⁹. Similarly, MHn (multi-scale Shannon entropy) extends classical information entropy by incorporating multiscale and nonlinear characteristics, thereby enhancing its ability to quantify signal complexity in time-series data. MHn follows the standard multiscale coarse-graining procedure, in which the original signal is averaged over non-overlapping windows at multiple temporal scales, and Shannon entropy is subsequently computed from the empirical distribution of the resulting coarse-grained signal at each scale^40,41. In total, 206 features were extracted from PA state sequences, including 4 single-scale features (SE, Hn, LZC, and activity percentage) and 202 multi-scale features (MSE and MHn computed across scales 1–100, as well as average MSE and average MHn). In addition, three demographic and psychological variables (CES-D⁴², age, and BMI) were included, resulting in a total of 209 features.

The following equations define the complexity measures used in our study. Specifically, Lempel-Ziv complexity (LZC), sample entropy (SE), information entropy (Hn), multi-scale sample entropy (MSE), and multi-scale information entropy (MHn) are calculated as follows:

$$\mathit{LZC}=\frac{c_{\alpha }\left(S\right(N,\alpha \left)\right)}{\frac{N}{{\log }_{\alpha }N}}$$ 3

where $N$ is the length of the sequence $S$, and $\alpha$ is the number of distinct symbols in the sequence.

$$\mathit{SE}\left(m,r,N\right)=-\ln \frac{A(m,r,N)}{B(m,r,N)}$$ 4

where $m$ is the length of the compared sequences, $r$ is the similarity threshold, $N$ is the number of data points, and function $A$ and $B$ is the average probability that two sequences of length $m$ are similar.

$$\mathit{Hn}\left(X\right)=-\sum _{x\in X}p\left(x\right)\mathit{logp}\left(x\right)$$ 5

where $X$ is the discrete random variable, and $x$ is a realization of $X$.

$$\mathit{MSE}\left(S_j,\tau ,r,K,L\right)=\frac{1}{K}\frac{\frac{1}{n-1}{\sum }_{k=1}^{n-{\tau }_i-1}{N}_{r,L}(S_{j,k}^{\left(i\right)},{\tau }_i)}{\frac{1}{n-1}{\sum }_{k=1}^{n-{\tau }_i-1}{N}_{r,L-1}(S_{j,k}^{\left(i\right)},{\tau }_i)}$$ 6

where $S_j$ is the time series segment at index $j$, $\tau$ is the time scale factor, $r$ is the similarity threshold, $K$ is the number of scale factors considered, and $L$ is the embedded dimension.

$$\mathit{MHn}\left(\tau \right)=-\sum _{k=1}^B{p}_k^{\left(\tau \right)}\mathit{log}\left(p_k^{\left(\tau \right)}\right),p_k^{\left(\tau \right)}=\frac{1}{M}\sum _{j=1}^{M}1\left(\frac{1}{\tau }\sum _{i=\left(j-1\right)\tau +1}^{j\tau }{x}_i\in {\mathcal{B}}_k\right)$$ 7

where $x_i$ is the original time-series value at index $i$, $\tau$ is the scale factor (window length), $j$ indexes the non-overlapping coarse-graining windows, $M=⌊N/\tau ⌋$ is the number of coarse-grained samples, ${\mathcal{B}}_k$ denotes the $k$-th bin (or state) used to estimate the empirical distribution, $p_k^{\left(\tau \right)}$ is the probability that the coarse-grained value at scale $\tau$ falls into bin $k$.

Normality was assessed using the Kolmogorov–Smirnov and Shapiro–Wilk tests. As most variables significantly deviated from a normal distribution (p < 0.05), non-parametric methods were employed for subsequent analyses. To identify features associated with physical frailty and concern about falling, group comparisons were performed using the Kruskal–Wallis test. Variables exhibiting statistical significance (p ≤ 0.05) in at least one of the two tasks, along with local features, were incorporated into the input set for machine learning classification.

To capture temporal dependencies and dynamic patterns in PA sequences, this study utilized bidirectional LSTM models, which are well-suited for handling vanishing gradients and retaining long-term information in sequential data. Physical activity (PA) sequences were first segmented into 48 h intervals. Within each interval, hourly distributions of PA states were analyzed, and the most frequently occurring PA state number was identified. Consequently, each sample was represented by a sequence summarizing its predominant activity status across each hour of the 48-hour window. This aggregation emphasizes stable, behaviorally meaningful activity states while reducing the influence of transient fluctuations. These condensed PA sequences were used as input for LSTM-based temporal neural networks, which effectively captured local temporal features by analyzing the progression and variation of activity states over time.

This study introduced a novel multi-task deep learning framework (Fig. 5), enhanced with attention mechanisms, to address the challenges of classifying individuals based on their levels of physical frailty and concern about falling. The framework was designed to simultaneously handle multiple tasks, leveraging shared and task-specific features to improve predictive performance. Physical frailty identification is framed as a three-class classification task, with labels: 0 (robust), 1 (pre-frail), and 2 (frail). Meanwhile, the identification of concern about falling is treated as a binary classification task, with subjects labeled as 1 (high concern) or 0 (low concern). Together, these define two related but distinct classification tasks, rather than a unified six-class grouping.

Fig. 5. Illustrated workflow of the multi-task learning model for joint prediction tasks.

Demographic, psychological, and global features are combined with local patterns extracted from PA state sequences using an LSTM-based network. The combined features are passed through an attention mechanism and linear layers to produce the final predictions.

The model input consisted of statistically significant features, including entropy-based activity complexity features, and demographic and psychological features, as well as 16-dimensional local embeddings extracted from PA sequences via bidirectional LSTM with normalization. Global features represent subject-level summary characteristics, whereas local features correspond to short-term temporal patterns encoded in the LSTM hidden representations. These features were concatenated to form a unified input vector and subsequently normalized to ensure consistent scale across inputs. The dimensionality, source, and role of each input feature group at the fusion layer are explicitly detailed in Table 7. The normalized vector was then passed into a customized, task-aware attention mechanism, which leveraged information from both tasks to enhance the feature weighting process through cross-task interaction^43,44. The resulting attention distributions were used to compute weighted representations of the input features, which were then fed into two independent fully connected linear layers⁴⁵, each responsible for predicting one of the target outcomes: physical frailty and concern about falling.

Table 7.

Input dimension explanation of the hybrid multi-task model

Stage	Features	Dimensions	Description
Feature Preparation	Demographic Features	(146, 2)	Age and body mass index (BMI); retained after statistical significance testing
	Psychological features	(146, 1)	CES-D score; retained after statistical significance testing
	Global handcrafted features	(146, 45)	Statistically significant global descriptors selected from the feature extraction process, including Hn, MHn, LZC, %activity, SE, MSE, MHn1–MHn38, and MHn41
	Physical activity (PA) states	(146, 48)	Hourly PA state sequences obtained by down-sampling per-second activity signals over a 48-hour window
Local Representation Learning	Local features (LSTM embeddings)	(146, 16)	Latent temporal embeddings learned by an LSTM from PA state sequences, capturing local and short-term activity dynamics
Feature Fusion & Attention	Concatenated feature vector	(146, 64)	Concatenation of statistically selected demographic (2), psychological (1), global (45), and learned local (16) features, used as input to the multi-task attention module

For the auxiliary task (concern about falling), an attention vector was first generated by applying a transformation and activation to the input features. This created a task-specific representation that highlighted important parts of the input. For the main task (physical frailty), a second attention mechanism combined the original input with the auxiliary task’s output. This helped the model focus on the most relevant information for the main task. By using the auxiliary task to guide the main task’s attention, the model benefits from shared learning and improved performance. The equations and detailed explanation of the attention mechanism are provided in Suppl. Table S7.

Since frailty classification involved three imbalanced classes, with the frail group comprising only 15.75% of the dataset, Focal Loss (Eq. 8) was employed to emphasize underrepresented classes and misclassified samples:

$$\mathit{FL}\left(p_t\right)=-{\alpha }_t{\left(1-p_t\right)}^{\gamma }\log \left(p_t\right)$$ 8

where ${\alpha }_t$ is a weighting factor to balance the importance of different classes, especially in the case of imbalanced datasets, $p_t$ is the predicted probability for the true class (i.e., the model’s confidence in the correct classification), $\gamma$ is the focusing parameter that reduces the loss contribution from well-classified examples, making the model focus more on hard-to-classify examples. Additionally, class weighting was incorporated into the Focal Loss function for the physical frailty task to ensure that the underrepresented yet clinically important frail class was sufficiently emphasized during training⁴⁶. These weights were automatically derived by computing values inversely proportional to the class frequencies, thereby assigning greater importance to less represented classes during training⁴⁷. Meanwhile, concern about falling, formulated as a binary classification task, was optimized using Binary Focal Loss (Eq. 9), which emphasizes hard-to-classify examples and mitigates class imbalance during training:

$$\mathit{BFL}\left(p_i\right)=-\frac{1}{N}\sum _{i=1}^N\alpha ·{\left(1-\widehat{p_i}\right)}^{\gamma }·y_i\log \left(\widehat{p_i}\right)+(1-\alpha )·{\widehat{p_i}}^{\gamma }·(1-y_i)\log (1-\widehat{p_i})$$ 9

where $N$ is the total number of samples, $y_i\in \left\{\mathrm{0,1}\right\}$ is the ground truth label for sample $i$, $\widehat{p_i}$ is the predicted probability obtained by applying the sigmoid function to the model output, $\alpha \in \left[\mathrm{0,1}\right]$ balances the importance between positive and negative samples, and $\gamma$ is the focusing parameter that down-weights easy examples and focuses training on hard negatives.

For comparison, several single-task learning models were implemented using the same experimental protocol. Classical machine learning approaches, including LR, SVM, and DT, were applied to non-sequential features comprising demographic, psychological, and global activity measures. In addition, sequence-based models were employed to capture temporal dependencies in PA data. Specifically, a bidirectional LSTM model and a TST were used to process PA state sequences. All single-task models were trained independently for each outcome, enabling a direct comparison with the proposed MTL framework.

To ensure the robustness and reliability of our models, we implemented stratified subject-wise k-fold cross-validation (CV). The dataset was split into 70% for training and 30% for testing. The training sets were further divided randomly into 5 folds. During each cycle, 4 folds were used for training, where the left fold was used for validation. For the multi-task learning framework, the final model was selected based on the highest average validation accuracy across both tasks, ensuring fair and balanced performance⁴⁸. After five training cycles, the best-performing model and its optimized hyperparameters—including learning rate, number of epochs, loss function, and architectural parameters (e.g., number of layers and neurons)—were retained. In the case of single-task learning models, the best model was identified separately for each task as the one achieving the highest validation accuracy after fine-tuning. The final selected models, which demonstrated the best overall validation performance, were then evaluated on the held-out testing set to generate predictions for both tasks. Each participant’s data was kept strictly independent across training and testing splits to avoid data leakage.

Author contributions

Jingyi Zhang led data processing, model development, manuscript writing, and prepared tables and figures. Jingtao Zhang contributed to manuscript revision and language editing. Peter Shull, Catherine Park, and Shaoxiong Sun provided critical feedback and contributed to manuscript revision. Bijan Najafi supported clinical interpretation. Changhong Wang was responsible for methodological design, manuscript writing, and overall project supervision. All authors reviewed and approved the final version of the manuscript.

Data availability

Due to ethical and privacy considerations, the datasets analyzed in this study are not publicly available. Researchers interested in accessing the data may contact the corresponding author for further information.

Code availability

All analyses were conducted in Python (version 3.12) using scikit-learn (version 1.5.2). Specific functions are detailed in the Methods section, with parameter settings provided in Suppl. Table S5. The deep learning framework code is publicly available at: https://github.com/JingyiZhangKTH/Frailty_Concern_About_falling_Prediction.

Competing interests

The authors declare no competing financial or non-financial interests. Author Peter Shull is Associate Editor of npj Digital Medicine. Peter Shull was not involved in the journal’s review of, or decisions related to, this manuscript.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41746-026-02863-4.